Gemini models increasingly stucking in thinking loop
Using Google Antigravity and Google AI Studio, I have noticed that Gemini models, which includes 3.1 Pro and 3.5 Flash, are increasingly stucking in thinking loops. It has become much more common. So I did a 100 tasks test and out of them, 3.5 Flash got stuck in a loop 23 times, while 3.1 Pro was a lower 16 times. Was this not the case earlier? I've had it stuck in loops from day 1, even when Gemini 3 came out. I've never been able to actually use it for anything because it goes absolutely mental after a handful of turns. It has become too common recently, even when the models used are same. I find when I use Gemini Web Chat that it often just starts over mid-conversation and doesn't remember the rest of the context of the conversation whereas Claude and ChatGPT never do that. All the tasks done were in Google Antigravity 2.0. I dunno about stuck, but I get a lot of errors that prevent me from submitting new prompts. Do you have a website available to show the sessions ? I have the website which I built in which it got stuck in a thinking loop. But it won't show any of the 100 tests I did. I am thinking of writing an article on this though. If you want to see, the site's Nasengetu. It is indexed so you can find it in Google search. conflicting instructions in prompts and though signatures being passed down across toolcall chains solved this for me Do you have a single example? I have multiple examples. Just yesterday, when I was trying to implement a caching system for my website Nasengetu, and it completely stuck in a loop-where it runs a log fetch using the git command and then thinks about the result, then runs the command again, and think again. And this repeated until I stopped it when it almost completely exhausted my five hour request limit. Getting stuck